Particle recognition and shape parameter detection based on deep learning

The size and shape parameters of sand particles are closely related to their geophysical and geomechanical properties. It is challenging to accurately identify sand particles and calculate their shape parameters. In this study, a convolutional neural network was used to detect sand particles in sample images and further calculate their size parameters. Using Mask R-CNN as the benchmark detection network, by analyzing the labeling data of sand particles, comparing the size of different a priori boxes to obtain better detection results. In addition, this study uses an edge detection algorithm with adaptive parameters to segment the particle region of interest, and combines the mask predicted by the network model to select the segmented region belonging to the object. The image processing algorithm can be used to segment the area more accurately, and the deep learning algorithm can detect the target more robustly, and combine the two to calculate the parameters of the particles.